Stochastic morphological evolution equations

Stochastic morphological evolution equations

J. Differential Equations 251 (2011) 2950–2979 Contents lists available at ScienceDirect Journal of Differential Equations www.elsevier.com/locate/j...

369KB Sizes 0 Downloads 141 Views

J. Differential Equations 251 (2011) 2950–2979

Contents lists available at ScienceDirect

Journal of Differential Equations www.elsevier.com/locate/jde

Stochastic morphological evolution equations ✩ Peter E. Kloeden ∗ , Thomas Lorenz Institut für Mathematik, Goethe-Universität, 60054 Frankfurt am Main, Germany

a r t i c l e

i n f o

a b s t r a c t

Article history: Received 14 February 2011 Available online 13 April 2011 MSC: primary 35R60, 35R70 secondary 34G25, 37H10, 49J53, 60H15, 93B03 Keywords: Nonlocal evolution equations Set-valued evolution equations Morphological equations Itô stochastic inclusion equations Random closed sets Nucleation

The inadequacy of locally defined set-valued differential equations to describe the evolution of shapes and morphological forms in biology, which are usually neither convex or nondecreasing, was recognised by J.-P. Aubin, who introduced morphological evolution equations, which are essentially nonlocally defined set-valued differential equations with the inclusion vector field also depending on the entire reachable set. This concept is extended here to the stochastic setting of set-valued Itô evolution equations in Hilbert spaces. Due to the nonanticipative nature of Itô calculus, the evolving reachable sets are nonanticipative nonempty closed random sets. The existence of solutions and their dependence on initial data are established. The latter requires the introduction of a time-oriented semi-metric in time-space variables. As a consequence the stochastic morphological evolution equations generate a deterministic nonautonomous dynamical system formulated as a two-parameter semigroup with the complication that the random subsets take values in different spaces at different time instances due to the nonanticipativity requirement. It is also shown how nucleation processes can be handled in this conceptual framework. © 2011 Elsevier Inc. All rights reserved.

1. Introduction The evolution of random sets, and of their shapes in particular, is of interest in a number of important applications. Traditional methods of set-valued analysis such as the Aumann integral are, however, restrictive and result in convex sets, while other assumptions result in nondecreasing sets, see [1,19,22]. Aubin [3,4] was well aware of these limitations from his investigations of determin✩ Partially supported by DFG grants KL 1203/7-1 and LO 273/5-1, the Ministerio de Ciencia e Innovación (Spain) grant MTM2008-00088 and the Junta de Andalucía grant P07-FQM-02468. Corresponding author. E-mail addresses: [email protected] (P.E. Kloeden), [email protected] (T. Lorenz).

*

0022-0396/$ – see front matter doi:10.1016/j.jde.2011.03.013

© 2011

Elsevier Inc. All rights reserved.

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2951

istic morphological evolution. Moreover he realised that set-valued processes generated by some set-valued differential equations, which are also known as differential inclusions or contingent equations, are inadequate for modelling morphological evolution because they are locally defined through the differential contingent, whereas nonlocal or even global dependence is required. Aubin [3,4] attempted to overcome this problem by introducing a class of nonlocally defined inclusion equations in which the contingent at each point in space-time, (t , x) ∈ [0, T ] × Rn , also depended on the entire set K (t ) ⊂ Rn attained at any time instant t ∈ [0, T ]. The resulting so-called morphological equations could be written ◦

K (t ) = F [t , K (t )] (t , x),

t ∈ [0, T ],

(1)

for a family of set-valued mappings F [t , M ] : [0, T ] × Rn ; Rn for each t ∈ [0, T ] and nonempty closed ◦

subset M of Rn . Here K (t ) represents the set of all elementary directions (vaguely, pseudo-derivatives) of the set-valued process K (t ) ⊂ Rn at time t. The nonlocal dependence comes from allowing the contingent set F [t , K (t )] (t , x) at each point (t , x) to depend also on the whole solution set K (t ) through the set-valued mappings F [t , K (t )] . This additional aspect opens the door to obtaining solutions whose set values need not be convex or increasing [3,4,20,21]. In this paper we develop stochastic counterparts of such morphological equations. They are to model the evolution of random set-valued phenomena in separable Hilbert spaces. In comparison with former approaches as in [1,22], we aim at significantly less constraints of the analytical concept. As in the deterministic case, the dynamics of our approach is not restricted to convex or expanding sets. This allows us to present a more realistic treatment of nucleation and birth-and-growth processes in a random environment. The only similar suggestion so far [21, § 3.7] is restricted to the Euclidean space, and it takes neither filtrations nor C 0 semigroups inducing mild solutions into consideration. Hence we can now cover a much broader class of stochastic set-valued models. Our counterparts are based on Itô stochastic differential inclusions, which are not differential equations at all, but integral equations based on Itô stochastic calculus. The analogy between the deterministic and stochastic cases is through the use of solutions to differential inclusions instead of selections of set-valued mappings. We will start with random closed sets generated by solutions to stochastic differential inclusions. They play the role of reachable sets, which are often called attainable sets, and extend the popular Aumann integral significantly. By means of standard results about stochastic differential inclusions, we establish existence results as well as useful estimates that we will then use in the corresponding stochastic generalisation of Aubin’s idea to what we call stochastic morphological evolution equations. The solution process is then single-valued process taking values in appropriate spaces (determined by nonanticipative conditions) of nonempty closed random sets. Considered as set-valued mappings it is indeed a set-valued stochastic process. On the other hand, considered as a single-valued mapping between spaces of nonempty closed random sets it defines a single-valued deterministic nonautonomous dynamical systems in the 2-parameter semigroup process formulation with the slight generalisation in that the state space can also change in time, see Kloeden and Rasmussen [18] as well as [15,16]. This is thus an alternative model of abstract random dynamical systems to the skewproduct flow like formalism of Arnold [2]. This paper is structured as follows. Section 2 provides the collection of relevant definitions and properties in regard to random closed sets, their mean square Hausdorff distance and Itô processes. In Section 3 we introduce reachable sets of stochastic differential inclusions and investigate their properties such as the continuous dependence on data. They serve as generalised integrators for the stochastic morphological evolution equations presented in Section 4. Theorem 4.2 is one of the main results. It provides sufficient conditions for the existence and uniqueness of solutions to stochastic morphological evolution equations. These solutions model shape evolutions with focus on growth and translation. Section 5 discusses their connection to dynamical systems in suitable abstract spaces briefly. Finally in Section 6 we suggest how a much broader class of initial random closed sets can be handled. Indeed the results before (like Theorem 4.2) assume the initial random closed set to be

2952

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

given at one single time instant. Appropriate approximation, however, enables us to cover a broad class of set-valued processes each of whose values initiates such a “growth” process (in the sense of Section 4). Expanding maps as in [1] prove to be a special case. Theorem 6.2 summarises our set-valued model with this additional aspect of birth. For the sake of a self-contained presentation, Appendix A provides the essential tool about stochastic differential inclusions here, i.e., Theorem A.1 of Da Prato and Frankowska, together with a proof. Topological properties of the space-time semimetric introduced in Section 2 is discussed in Appendix B. 2. Random closed sets Let (Ω, A, P) be a complete probability space. Let H be a separable real Hilbert space and let B denote the Borel σ -algebra on H . Definition 2.1. A random closed set in H is a measurable set-valued map (Ω, A) ; ( H , B ) for which the values are nonempty closed subsets of H . The set of all these random closed sets is denoted by RC (Ω, A; H ). A random closed set M ∈ RC (Ω, A; H ) is called square integrable if there exists at least one square integrable selection f : Ω → H of M, i.e., f ∈ L 2 (Ω, A, P; H ) with f (ω) ∈ M (ω) for all ω ∈ Ω . 2 ( M ) ⊂ L 2 (Ω, A, P; H ) denote the set of all square integrable selections Ω → H of M and Let SRC

let RC 2 (Ω, A; H ) denote the set of square integrable random closed sets Ω ; H . A random closed set M ∈ RC 2 (Ω, A; H ) is called square integrably bounded if

 Def.  E  M 2∞ =

 sup | z|2 dP(ω) < ∞.

Ω

z ∈ M (ω )

2 For every M ∈ RC 2 (Ω, A; H ), SRC ( M ) is closed in L 2 (Ω, A, P; H ). Hence each separable closed subset of L 2 (Ω, A, P; H ) induces a square integrable random closed set uniquely (see also Proposition 2.4 below). A square integrable random closed set M ∈ RC 2 (Ω, A; H ), however, does not have to be square integrably bounded. By Castaing’s Characterisation Theorem [7, Theorem 8.1.4], a set-valued map M : (Ω, A) ; ( H , B ) with nonempty closed values is measurable if and only if it is the pointwise closure of the union of (at most) countably many measurable functions f n : Ω → H , n ∈ N, i.e.,



M (ω) =

f n (ω)

for every ω ∈ Ω.

n∈N

Hence every square integrable random closed set M : Ω ; H can be represented by a sequence ( gn )n∈N in L 2 (Ω, A, P; H ) as



M (ω) =

gn (ω)

for every ω ∈ Ω.

n∈N

Indeed, there exists at least one selection g ∈ L 2 (Ω, A, P; H ) of M by definition and for each m ∈ N and f n the auxiliary function gn,m : Ω → H defined by

 gn,m :=

fn g

if m − 1  | f n | < m, otherwise

is square integrable. The importance of the following statement quoting [25, Lemma 2.1.3] (for p = 2) is that it allows many conclusions about any set M ∈ RC 2 (Ω, A; H ) to be obtained from just a single pointwise covering family ( gn )n∈N in L 2 (Ω, A, P; H ).

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2953

Lemma 2.2 (Approximation by step-functions). Let the sequence ( gn )n∈N in L 2 (Ω, A, P; H ) satisfy M (ω) = n∈N gn (ω ) for any ω ∈ Ω . 2 Then, for every f ∈ SRC ( M ) and ε > 0, there exists a finite measurable partition A 1 , A 2 , . . . , Am of Ω such that



  m     g j · χA j  f −   j =1

< ε. L 2 (Ω)

The general assumption P(Ω) = 1 and the standard selection theorem about marginal maps (see e.g. Theorem 8.2.11 of [7]) now lead to this equivalence: Corollary 2.3. The following properties are equivalent for any M ∈ RC 2 (Ω, A; H ) and c  0: (1) M is square integrably bounded, i.e., E( M 2∞ )  c 2 < ∞. 2 (2) SRC ( M ) is bounded in L 2 (Ω, A, P; H ) by c. (3) Some sequence ( gn )n∈N in L 2 (Ω, A, P; H ) with M (ω) =



n∈N

gn (ω) for any ω ∈ Ω satisfies

sup  gn  L 2 (Ω; H )  c < ∞. n∈N

The following general criterion for subsets of L 2 (Ω, A, P; H ) to induce random closed sets is taken from Molchanov [25, Theorem 2.1.6]. Proposition 2.4 (Decomposable sets and selections). Let Ξ be a nonempty closed subset of L 2 (Ω, A, P; H ) and let χ B : Ω → {0, 1} denote the characteristic function of a subset B of Ω . 2 Then, there exists a random closed set M : Ω ; H with SRC ( M ) = Ξ if and only if Ξ is decomposable in the following sense: For any ξ1 , ξ2 ∈ Ξ and B ∈ A, the function χ B · ξ1 + (1 − χ B ) · ξ2 : Ω → H also belongs to Ξ . It provides the following implication when the nonempty set Ξ ⊂ L 2 (Ω, A, P; H ) is not assumed to be closed. Corollary 2.5. Let Ξ be a nonempty subset of L 2 (Ω, A, P; H ). If Ξ is decomposable, then there exists a square 2 ( M ) of L 2 selections is equal to the closure of Ξ . integrable random closed set M : Ω ; H for which the set SRC 2.1. Adapted selections of set-valued maps Let (Ω, A, {At }t 0 , P) be a complete filtered probability space satisfying the usual hypothesis, i.e., {At }t 0 is an increasing and right continuous family of σ -sub-algebras of A and A0 contains all P-null sets. Denote by PA the predictable σ -algebra for the filtration {At }t 0 , i.e., the σ -algebra on [0, ∞) × Ω generated by all sets of the form {0} × A (for any A ∈ A0 ) and (s, t ] × A (with any 0  s < t < ∞ and A ∈ A s ). Definition 2.6. Let ( X , d) be a complete separable metric space and let J ⊂ [0, ∞) be measurable. A set-valued map F : J × Ω ; X is called adapted with respect to a filtration {At }t 0 if the set-valued map F (t , ·) : Ω ; X is At -measurable for each t ∈ J , i.e., the inverse image of each open set O ⊂ X is At -measurable:

F (t , ·)−1 ( O ) =

Def.







ω ∈ Ω F (t , ω) ∩ O = ∅ ∈ At .

2954

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

A set-valued map F : J × Ω ; X is called predictable if it is measurable with respect to PA , i.e., the inverse image of each open set O ⊂ X is PA -measurable:

F −1 ( O ) =

Def.



(t , ω) ∈ J × Ω F (t , ω) ∩ O = ∅ ∈ PA .

A close inspection of the proofs in [7, § 8] shows that many standard results about measurable selections can be extended to adapted and predictable selections, respectively. Proposition 2.7 (Adapted and predictable selections of set-valued maps). Suppose the set-valued map F : J × Ω ; X to be adapted with respect to {At }t 0 and to have nonempty closed values. Then, there exists a selection of F adapted with respect to {At }t 0 , i.e., an adapted single-valued function f : J × Ω → X satisfying f (t , ω) ∈ F (t , ω) for every t ∈ J and ω ∈ Ω . If F : J × Ω ; X is predictable, then so is the selection f : J × Ω → X . Theorem 2.8 (Characterisation of predictable set-valued maps). For any set-valued map F : J × Ω ; X with nonempty closed values, the following properties are equivalent: (a) F is predictable. (b) For x ∈ X , the function dist(x, F (·,·)) : J × Ω → [0, ∞) is PA -measurable. (c) There exists a sequence ( f n )n∈N of predictable selections of F such that

F (t , ω) =



f n (t , ω)

for every (t , ω) ∈ J × Ω.

n∈N

It follows that every predictable set-valued map is both A-measurable and adapted with respect to {At }t 0 . 2.2. The mean square Hausdorff distance The mean square Hausdorff excess (or semi-distance) between random closed sets in H is defined as

e ⊂ RC ( M 1 , M 2 ) :=

sup 2 f ∈SRC (M1 )

  E dist( f , M 2 )2 ,

and the mean square Hausdorff distance as





⊂ dlRC ( M 1 , M 2 ) := max e ⊂ RC ( M 1 , M 2 ), e RC ( M 2 , M 1 ) .

They may be infinite valued. By a general result on the Hausdorff metric topology of nonempty closed subsets in any complete metric space [8, Theorem 3.2.4 (1.)], [10, Theorem II.3], it follows from the completeness of L 2 (Ω, A, P; H ) that: Lemma 2.9. RC 2 (Ω, A; H ) is complete with respect to dlRC . In combination with the well-known Selection Theorem of Kuratowski and Ryll-Nardzewski, Lemma 2.2 ensures for any sets M 1 , M 2 ∈ RC 2 (Ω, A; H ) and a sequence ( f n )n∈N in L 2 (Ω, A, P; H )  with M 1 (ω) = n∈N f n (ω) for every ω ∈ Ω that

sup 2 f ∈SRC (M1 )

  E dist( f , M 2 )2 =

sup

inf

2 f ∈SRC ( M 1 ) g ∈SRC ( M 2 )

= sup

2

inf

2 (M2 ) m∈N g ∈SRC

  E | f − g |2

 f m − g 2L 2 (Ω; H ) .

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2955

As a consequence,

dlRC ( M 1 , M 2 ) = max

 sup

 = sup

inf

2 f ∈SRC ( M 1 ) g ∈SRC ( M 2 ) 2

inf

2 f ∈SRC (M1 )

 f − g L 2 (Ω; H ) ,

h − f L 2 −

inf

2 g ∈SRC (M2 )

sup

inf

2 2 g ∈SRC ( M 2 ) f ∈SRC ( M 1 )

 f − g L 2 (Ω; H )



 h − g L 2 h ∈ L 2 (Ω, A, P; H )

is a metric on RC 2 (Ω, A; H ), which may take infinite values. Moreover, for any three sets M 1 , M 1 , M 2 ∈ RC 2 (Ω, A; H )









⊂ ⊂ e ⊂ RC M 1 ∪ M 1 , M 2 = max e RC ( M 1 , M 2 ), e RC M 1 , M 2



.

2.3. Itô processes Let H and U be separable real Hilbert spaces. Let Lin(U , H ) consist of all bounded linear operators U → H and define  M ∞ := sup y ∈ M | y | H , which may be infinite, for any subset M ⊂ H . (Henceforth, the norms | · | H and | · |U will usually be written simply as | · |.) Suppose A to be the generator of a strongly continuous semigroup ( S (t ))t 0 of linear bounded operators on H . In addition, ( S (t ))t 0 is assumed to be ρ -contractive, i.e., there exists some ρ > 0 with

S (t )x  e ρ t |x| for all t  0 and x ∈ H . For a fixed finite T > 0 let L2A ([0, T ], H ) be the class of all (single-valued) functions f : [0, T ] × Ω → H such that f is predictable (i.e., PA -measurable, so f (t , ·) : Ω → H is At -measurable for every t ∈ [0, T ]), and satisfies



2   E f (t , ·) dt < ∞,

2   E f (t , ·) < ∞

[0, T ]

for every t ∈ [0, T ]. Now let W = ( W t )t 0 be a Wiener process in (Ω, A, P) that is adapted to {At }t 0 with values in U . Suppose Q to be a positive finite trace operator in U such that for all t  0 and u ∈ U ,

2   E  W t , u U = t ·  Q u , u U . As summarised in [13], for example, the Itô integral valued predictable stochastic process Φ with

T



T 0

Φ(s) dW s is well defined for any Lin(U , H )-



trace Φ(s)Φ ∗ (s) Q ds < ∞ P-almost surely.

0

T

If, moreover, E(

0

trace(Φ(s)Φ ∗ (s) Q ) ds) < ∞, then the following identity (Itô isometry) holds

2   T  T    ∗ E Φ(s) dW s = E trace Φ(s)Φ (s) Q ds . 0

0

2956

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Furthermore, there exists a constant C = C ( T ) with

 t 2   t   2   Φ(s) Q ds E sup S (t − s)Φ(s) dW s  C · E s∈[0,t ] 

0

0



with the operator norm defined by  S  Q := trace( S S ∗ Q ) for any S ∈ Lin(U , H ). Let I0 (t 0 , X 0 , γ , σ ) denote the Itô process in H with the initial time t 0 ∈ [0, T ], initial state X 0 ∈ L 2 (Ω, At0 , P; H ), drift γ ∈ L2A ([t 0 , T ], H ) and diffusion σ ∈ L2A ([t 0 , T ], Lin(U , H )), i.e.,

t I0 (t 0 , X 0 , γ , σ )t := S (t ) X 0 +

t S (t − s)γ (s) ds +

t0

S (t − s)σ (s) dW s t0

for t ∈ [t 0 , T ] and define

  I0 (t 0 , X 0 , γ , σ )

I,[t 0

   t   t      2 := E | X 0 |2 + E |γ |2 ds + E σ  Q ds . ,t ] t0

t0

An Itô process satisfies the general estimate

2   2  E I0 (t 0 , X 0 , γ , σ )t  4(1 + t − t 0 )e 2ρ (t −t0 ) I0 (t 0 , X 0 , γ , σ )I,[t ,t ] 0  2 (1+2ρ )(t −t 0 )    4e I0 (t 0 , X 0 , γ , σ ) I,[t ,t ] , 0

(2)

which follows from the Hölder inequality, the Itô isometry and the simple inequality (r + s)2  2(r 2 + s2 ) for any r , s ∈ R. 2.4. A time-oriented distance for handling the C 0 semigroup in time The presence of adaptivity suggests that the time variable should be included in the state of the system, i.e., (t , X t ) instead of just X t ∈ L 2 (Ω, At , P; H ). This simple supplement facilitates handling continuity of Itô processes with respect to time in the following way. As in Section 2.3 assume U , H to be real separable Hilbert spaces and ( S (t ))t 0 to be a ρ -contractive strongly continuous semigroup of linear bounded operators on H . Now consider the basic set of states



E L := (t , X ) t ∈ [0, T ], X ∈ L 2 (Ω, At , P; H ) and define



L : E L × E L → [0, ∞[,   |s − t | + E(| S (t − s) X − Y |2 ) if s  t , (s, X ), (t , Y ) → |s − t | + E(| X − S (s − t )Y |2 ) if s > t .

Obviously L is positive definite and symmetric. Moreover it satisfies the following time-oriented modification of the triangle inequality

      L (t 1 , X 1 ), (t 3 , X 3 )  2e 2ρ (t3 −t2 ) · L (t 1 , X 1 ), (t 2 , X 2 ) + 2 · L (t 2 , X 2 ), (t 3 , X 3 )

(3)

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2957

for any (t 1 , X 1 ), (t 2 , X 2 ), (t 3 , X 3 ) ∈ E L with t 1  t 2  t 3 because

2     L (t 1 , X 1 ), (t 3 , X 3 ) = t 3 − t 1 + E S (t 3 − t 1 ) X 1 − X 3 2      t 3 − t 1 + E S (t 3 − t 2 ) S (t 2 − t 1 ) X 1 − X 2 + S (t 3 − t 2 ) X 2 − X 3 2   2    t 3 − t 1 + 2 · E S (t 3 − t 2 ) S (t 2 − t 1 ) X 1 − X 2 + S (t 3 − t 2 ) X 2 − X 3 2 2    t 3 − t 1 + 2 · E e 2ρ (t3 −t2 ) S (t 2 − t 1 ) X 1 − X 2 + S (t 3 − t 2 ) X 2 − X 3 . However, L is not a metric in the classical sense, but nevertheless leads to the same concept of sequential convergence as the norm of [0, T ] × L 2 (Ω, A, P; H ). Indeed, ( S (t ))t 0 induces a strongly continuous semigroup on L 2 (Ω, A, P; H ) and so one concludes from [21, Proposition 3.117] that for any bounded sequence ((tk , X k ))k∈N in E L and (t , X ) ∈ E L ,





lim L (tk , X k ), (t , X ) = 0

k→∞

⇐⇒





lim |tk − t | + E | X k − X |2

k→∞



= 0.

(4)

In comparison with the norm, however, the key advantage of L is that all curves [0, T ] → E L , t → S (t ) X 0 with any X 0 ∈ L 2 (Ω, A0 , P; H ) are 1-Lipschitz continuous. A similar equi-continuity with respect to the norm is known not to hold even locally, unless the generator A of the semigroup is bounded linear. This observation implies for the Itô process X := I0 (t 0 , X 0 , γ , σ ) specified in Section 2.3 and any t0  s  t  T

 t    t   2ρ (t −s) 2 2 L (s, X s ), (t , X t )  |t − s| + 2e |γ | dr + E σ  Q dr (t − s) · E 

s

 2  |t − s| + 2e (1+2ρ )(t −s) I0 (t 0 , 0, γ , σ )I,[s,t ] .

s

(5)

See Appendix B for the counterpart RC concerning square integrable random closed sets and its topological properties. 3. Reachable sets of stochastic differential inclusions The concept of a reachable set is extended here to stochastic differential inclusions in a separable real Hilbert space H . Let W = ( W t )t 0 be a fixed Wiener process with values in a separable Hilbert space U and adapted to a filtration {At }t 0 for the underlying probability space (Ω, A, P). Specifically, the definition of “reachable set” uses square integrable strong solutions expressed in terms of Itô processes. A generalisation of the Filippov-like theorem of Da Prato and Frankowska (see Theorem A.1) is the key tool for investigating its dependence on the given data with respect to the Hausdorff distance dlRC . Anticipating the further generalisation to stochastic morphological evolution equations, an initial closed random set rather than a single initial random variable will be considered. Definition 3.1 (Random reachable set). Consider a set-valued map F = ( F 1 , F 2 ) : [0, T ] × Ω × H ; H × Lin(U , H ). The random reachable set ϑ F (t ; t 0 , K 0 ) : Ω ; H of the stochastic differential inclusion





d X t ∈ A X t + F 1 (t , X t ) dt + F 2 (t , X t ) dW t and initial state (t 0 , K 0 ) ∈ [0, T ] × RC 2 (Ω, At0 ; H ) at time t ∈ [t 0 , T ] is defined in terms of strong Itô solutions as

2958

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

 

ϑ F (t ; t 0 , K 0 )(ω) := X t (ω) ∃ X 0 ∈ L 2 (Ω, At0 , P; H ), γ ∈ L2A [t 0 , T ], H ,   σ ∈ L2A [t 0 , T ], Lin(U , H ) : Xt = I0 (t 0 , X 0 , γ , σ )t with X 0 ∈ K 0 ,     P-a.s., and γ (s, ω) ∈ F 1 s, ω, X s (ω) , σ (s, ω) ∈ F 2 s, ω, X s (ω)   for L1 × P -a.a. (s, ω) ∈ [t 0 , t ] × Ω . The random closed reachable set ϑ F (t ; t 0 , K 0 ) : Ω ; H is the random closed set for which the set of selections in L 2 (Ω, At , P; H ) coincides with the closure of all these solutions X t ∈ L 2 (Ω, At , P; H ) starting in K 0 , P-a.s. 2 Obviously, ϑ F (t 0 ; t 0 , K 0 ) = K 0 holds for every K 0 ∈ RC (Ω, At 0 ; H ) and any set-valued map  F : [0, T ] × H ; H × Lin(U , H ) because K 0 is characterised by S 2 ( K 0 ) = ∅.

RC

2 The next step is to verify that ϑ F (t ; t 0 , K 0 ) is well defined in RC (Ω, At ; H ). This follows from Proposition 2.4 on decomposable sets and selections and Corollary 2.5. It is due, essentially, to the pathwise structure of the stochastic differential inclusion.

F : [0, T ] × Ω × H ; H × Lin(U , H ) with nonempty closed values, K 0 ∈ Proposition 3.2. For every  RC 2 (Ω, At0 ; H ) and t ∈ [t 0 , T ], the set-valued map ϑF (t ; t 0 , K 0 ) : Ω ; H defined via the L 2 closure of solutions X t ∈ L 2 (Ω, At , P; H ) starting at K 0 is a square integrable random closed set, i.e., ϑ F (t ; t 0 , K 0 ) ∈ RC 2 (Ω, At ; H ). Proof. At a given time t ∈ [t 0 , T ] consider the set Ξt of all X t ∈ L 2 (Ω, At , P; H ) induced by an Itô process X = I0 (t 0 , X 0 , γ , σ ) that solves the stochastic differential inclusion





d Xs ∈ A Xs +  F 1 (s, X s ) ds +  F 2 (s, X s ) dW s 2 and starts in K 0 ∈ RC 2 (Ω, A0 ; H ), i.e., X 0 ∈ SRC ( K 0 ) and





γ (s, ω ) ∈  F 1 s, ω, X s (ω) ,





σ (s, ω ) ∈  F 2 s, ω, X s (ω)

for (L1 × P)-almost all (s, ω) ∈ [t 0 , t ] × Ω . It will be shown that this subset Ξt is decomposable, because then Corollary 2.5 provides a set 2 K t ∈ RC 2 (Ω, At ; H ) with SRC ( K t ) = Ξt ⊂ L 2 (Ω, At , P; H ) and by definition, ϑF (t ; t 0 , K 0 ) = K t . For this, choose B ∈ A, X t ∈ Ξt and  X t ∈ Ξt arbitrarily. Then, by definition of Ξt , there 2 X 0 ∈ SRC ( K 0 ), γ , γ˜ ∈ L2A ([t 0 , T ], H ) and σ , σ˜ ∈ L2A ([t 0 , T ], Lin(U , H )) such that X t = exist X 0 ,  I0 (t 0 , X 0 , γ , σ )t and  X t = I0 (t 0 ,  X 0 , γ˜ , σ˜ )t . Write





γ := χ B · γ + (1 − χ B ) · γ˜ ∈ L2A [t 0 , T ], H ,

   σ := χ B · σ + (1 − χ B ) · σ˜ ∈ L2A [t 0 , T ], Lin(U , H ) , 2  X 0 := χ B · X 0 + (1 − χ B ) ·  X 0 ∈ SRC ( K 0 ),

,  and let  X t := I0 (t 0 ,  X0, γ σ )t . Then,



τ S (τ − s) γ (s) ds = χ B ·

t0

τ S (τ − s)γ˜ (s) ds

S (τ − s)γ (s) ds + (1 − χ B ) · t0

t0

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

holds everywhere in Ω for every

τ ∈ [t 0 , t ] and the corresponding statement about Itô integrals, i.e.,



τ S (τ − s) σ (s) dW s = χ B ·

t0

2959

τ S (τ − s)σ˜ (s) dW s

S (τ − s)σ (s) dW s + (1 − χ B ) · t0

t0

also holds P-almost surely in Ω . Hence,

 X τ = χ B · X τ + (1 − χ B ) ·  Xτ , P-almost surely in Ω , for every τ ∈ [t 0 , t ]. F 1 (s, ω, X s (ω)), σ (s, ω) ∈  F 2 (s, ω, X s (ω)) and γ˜ (s, ω) ∈ Furthermore, the inclusions γ (s, ω) ∈   F 1 (s, ω ,  X s (ω)), σ˜ (s, ω) ∈  F 2 (s, ω ,  X s (ω)) are satisfied for (L1 × P)-almost all (s, ω) ∈ [t 0 , t ] × Ω by definition of Ξt . Finally, one obtains













γ(s, ω) ∈  F 1 s, ω, χ B (ω) · X s (ω) + 1 − χ B (ω) ·  X s (ω) =  F 1 s, ω ,  X s (ω) ,

       σ (s, ω ) ∈  F 2 s, ω, χ B (ω) · X s (ω) + 1 − χ B (ω) ·  X s (ω) =  F 2 s, ω ,  X s (ω) , for (L1 × P)-almost all (s, ω) ∈ [t 0 , t ] × Ω , i.e.,

χ B · Xt + (1 − χ B ) ·  X t ∈ Ξt ⊂ L 2 (Ω, At , P; H ). 2

It is not clear how to ensure the closedness of Ξt ⊂ L 2 (Ω, At , P; H ) without very restrictive 2 ( K 0 ), even if H is finite-dimensional. Thus the L 2 closure of selections is assumptions about SRC preferred here as an additional step in the construction of ϑ F (t ; t 0 , K 0 ).

Definition 3.3. For Λ > 0 fixed, denote by PLIPΛ ([0, T ] × Ω × H ; H × Lin(U , H )) the class of all setvalued maps F = ( F 1 , F 2 ) : [0, T ] × Ω × H ; H × Lin(U , H ) with the following properties: (1) (2) (3) (4)

F has nonempty bounded closed values, F (·,·, x) is predictable for every x ∈ H , F (t , ω, ·) is Λ-Lipschitz continuous for every t ∈ [0, T ] and F (·,·, 0) is uniformly bounded, i.e.,

ω ∈ Ω,



 Def.  F lg = sup  F (t , ω, 0)∞ t ∈ [0, T ], ω ∈ Ω < ∞. The following estimates show that the set-valued mapping

ϑF :

 

 [t 0 , T ] × {t 0 } × RC 2 (Ω, At0 ; H ) → RC 2 (Ω, A; H )

t 0 ∈[0, T ]

is Lipschitz continuous in time – but with respect to the square distance dlRC (·,·)2 rather than the metric dlRC itself. Proposition 3.4. Suppose that F , G ∈ PLIPΛ ([0, T ] × Ω × H ; H × Lin(U , H )) for some Λ > 0. Then, for any 0  t 0  s  t and K 0 ∈ RC 2 (Ω, At0 ; H ),

2       E ϑ F (t ; t 0 , K 0 )∞  e const(Λ,ρ ,T ) E  K 0 2∞ + 2 F 2lg t ,         RC s, ϑ F (s; t 0 , K 0 ) , t , ϑ F (t ; t 0 , K 0 )  (t − s) · const(Λ, ρ , T ) E  K 0 2∞ +  F lg

(6) (7)

2960

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

and if, in addition, F is globally bounded, i.e.,  F ∞ := supt ,ω,x  F (t , ω, x)∞ < ∞, then

      RC s, ϑ F (s; t 0 , K 0 ) , t , ϑ F (t ; t 0 , K 0 )  const(ρ , T ) 1 +  F ∞ · (t − s).

(8)

Moreover, for any K 1 , K 2 ∈ RC 2 (Ω, At0 ; H ),



2

dlRC ϑ F (t ; t 0 , K 1 ), ϑG (t ; t 0 , K 2 )



t 2

 C · dlRC ( K 1 , K 2 ) +

sup dlRC

t0

Y ∈L2



 2 C (1+t −t 0 ) F (s, ·, Y ), G (s, ·, Y ) ds · e (t −t0 )·e

(9)

with a constant C  1 that depends only on Λ and ρ . Proof. Proof of (6). We assume in addition that K 0 is square integrably bounded, i.e., E( K 0 2∞ ) < ∞, because otherwise the claim is obvious. For any given initial time t 0 ∈ [0, T ], initial random variable X 0 ∈ L 2 (Ω, At0 , P; H ), drift γ ∈ L2A ([t 0 , T ], H ) and diffusion σ ∈ L2A ([t 0 , T ], Lin(U , H )), inequality (2) provides a general estimate for the Itô process X = I0 (t 0 , X 0 , γ , σ ) at time t ∈ [t 0 , T ], namely

 2   E | X t |2  4e (1+2ρ )(t −t0 ) I0 (t 0 , X 0 , γ , σ )I,[t  Def.

= 4e



(1+2ρ )(t −t 0 )

E | X0|

2



 t

0 ,t ]



 t

2

+E

|γ | ds + E



t0

  4e



(1+2ρ )(t −t 0 )

E | X0|

2



 t +E

 2Q

ds

t0





2

 F lg + Λ| X s | ds

t0

  4e



(1+2ρ )(t −t 0 )

E | X0|

2



 t +E



2 F 2lg

2

2





+ 2Λ | X s | ds

t0

 = 4e



(1+2ρ )(t −t 0 )

E | X0|

2



 t + 2 F 2lg (t

2

 2

− t 0 ) + 2Λ · E

| X s | ds t0

 = 4e

(1+2ρ )(t −t 0 )



E | X0|

2



t + 2 F 2lg (t

2

− t 0 ) + 2Λ ·



2





E | X s | ds . t0

Hence, by Gronwall’s inequality,

2       E I0 (t 0 , X 0 , γ , σ )t  4e (1+2ρ )(t −t0 ) E | X 0 |2 + 2 F 2lg (t − t 0 ) · e const(Λ,ρ ,T )·(t −t0 )      e const(Λ,ρ ,T ) E | X 0 |2 + 2 F 2lg t . This implies for every {At }-adapted solution X = I0 (t 0 , X 0 , γ , σ ) : [t 0 , T ] → L 2 (Ω, A, P; H ) to the stochastic differential inclusion





d X t ∈ A X t + F 1 (t , X t ) dt + F 2 (t , X t ) dW t

(10)

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2961

starting in X 0 at time t 0 that

2         E | X t |2 = E I0 (s, X s , γ , σ )t  e const(Λ,ρ ,T ) E  K 0 2∞ + 2 F 2lg t . Hence the claimed estimate (6) results from Corollary 2.3 and Definition 3.1 of ϑ F (t ; t 0 , K 0 ). Proof of (7). Inequality (4) states for any Itô process X := I0 (t 0 , X 0 , γ , σ ) and t 0  s  t  T

 2   L (s, X s ), (t , X t )  |t − s| + 2e (1+2ρ )(t −s) I0 (t 0 , 0, γ , σ )I,[s,t ] . If, in addition, X is a solution to the stochastic differential inclusion (10), then one obtains exactly as in the preceding step

  t    (1+2ρ )(t −s) 2 2 2 L (s, X s ), (t , X t )  t − s + 2e · 2  F lg (t − s) + Λ · E | X r | dr 

s

(6 )

 t − s + 4e (1+2ρ )(t −s) · (t − s)      ×  F 2lg + Λ2 · e const(Λ,ρ ,T ) E  K 0 2∞ + 2 F 2lg t      const(Λ, ρ , T ) · E  K 0 2∞ +  F 2lg · (t − s). Def.

Proof of (8). Under the additional assumption  F ∞ = supt ,ω,x  F (t , ω, x)∞ < ∞, it follows similarly for the Itô process X t = I0 (t 0 , X 0 , γ , σ )t at time t ∈ [t 0 , T ] that for any s ∈ [t 0 , t ],

 t    t   2ρ (t −s) 2 2 L (s, X s ), (t , X t )  t − s + 2e |γ | dr + E σ  Q dr (t − s)E 

2ρ (t −s) t −s



s 2

2



s

 F 1 ∞ +  F 2 ∞ (t − s) e    const(ρ , T ) 1 +  F ∞ · (t − s).

 t − s + 2e

Proof of (9). Due to the symmetry of dlRC and the definition of ϑ F (t ; t 0 , K 1 ) through the L 2 closure, it is sufficient to prove for every solution X = I0 (t 0 , X 0 , γ , σ ) : [t 0 , t ] → L 2 (Ω, A, P; H ) to the stochastic inclusion (10) starting at a selection X 0 ∈ L 2 (Ω, At0 , P; H ) of K 1 ∈ RC 2 (Ω, At0 ; H ) that

  2  E dist X t , ϑG (t ; t 0 , K 2 )   t   2  C (1+t −t 0 ) 2  C E dist( X 0 , K 2 ) + dlRC F (s, X s ), G (s, X s ) ds · e (t −t0 )·e t0

with a constant C  1 which depends only on Λ and ρ . 2 Choose a selection Y 0 ∈ SRC ( K 2 ) ⊂ L 2 (Ω, At0 , P; H ) with  X 0 − Y 0 2L 2 = E(dist( X 0 , K 2 )2 ) by means of Theorem 8.2.11 of [7] (about measurable marginal maps) and the well-known Selection Theorem of Kuratowski and Ryll-Nardzewski. According to Theorem A.1 of Da Prato and Frankowska (see Appendix A), there exist a constant c = c (Λ, ρ )  1 and a strong solution Y : [t 0 , t ] → L 2 (Ω, A, P; H ) of





dY s ∈ AY s + G 1 (s, Y s ) ds + G 2 (s, Y s ) dW s

2962

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

starting in Y 0 such that

 2



Y − X I,[t0 ,t ]  c · E |Y 0 − X 0 |

2



t +

  2  c (1+t −t 0 ) E dist γ (s), σ (s) , G (s, X s ) ds · e (t −t0 )·e . 



t0

Then, inequality (2) and the definition of dlRC (·,·)2 through L 2 selections imply that

    2  E dist X t , ϑG (t , K 2 )  E |Y t − X t |2 (1+2ρ )(t −t 0 )

 4e

t +

dlRC







c · E dist( X 0 , K 2 )2



 2 c (1+t −t 0 ) F (s, X s ), G (s, X s ) ds · e (t −t0 )·e .

2

t0

Remark 3.5. The upper bound (9) in Proposition 3.4 takes

2 Def. =



dl∞ F (s, ·,·), G (s, ·,·)

2



dlRC F (s, ·, Y ), G (s, ·, Y )

sup Y ∈ L 2 (Ω,A

s ,P; H )

into consideration and, for later, it is recommendable to ensure that this distance is integrably bounded in time. According to Definition 3.3, however, the set-valued mappings F (s, ω, ·), G (s, ω, ·) : H ; H × Lin(U , H ) are Λ-Lipschitz continuous for all s ∈ [0, T ] and ω ∈ Ω . This motivates the alternative distance



2 Def. =

dllg F (s, ·,·), G (s, ·,·)

sup Y ∈ L 2 (Ω,As ,P; H )

dlRC ( F (s, ·, Y ), G (s, ·, Y ))2 1 + |Y |2

between the coefficients of differential inclusions. The latter is finite for a much broader class of setvalued maps in PLIPΛ ([0, T ] × Ω × H ; H × Lin(U , H )) and the estimate (9) can be adapted easily. Indeed, the estimate (6) of Proposition 3.4 provides an a priori bound of solutions to the corresponding stochastic inclusions that can occur in the proof. The consequences on estimate (9) will be reflected by an additional factor depending on E( K 1 2∞ ),  F lg , G lg , Λ, ρ and T . For the sake of the transparency, dl∞ is preferred here and the generalisation using dllg is left to the reader. 4. Stochastic morphological evolution equations Stochastic differential equations are, in fact, stochastic integral equations and are only written symbolically as differential equations. Nevertheless, the existence theory of such equations has many common features with that of ordinary differential equations written in their equivalent integral equation form. This is equally valid for both the single and set-valued cases. Here we obtain a stochastic generalisation of Aubin’s morphological equations by allowing the setvalued coefficients of stochastic differential inclusions to be modified by their reachable sets, which are described by square integrable random closed sets. This is the essential new aspect compared with earlier approaches in [1,22]. To be able to handle the adaptivity of random closed sets, the basic set of states is now



ERC := (t , M ) t ∈ [0, T ], M ∈ RC 2 (Ω, At ; H )

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2963

(as mentioned in Section 2.4) while the drift and diffusion coefficients, which now depend on such states, are given by single-valued function F with

  F[t , M ] ∈ PLIPΛ {t } × Ω × H ; H × Lin(U , H ) for every (t , M ) ∈ ERC . Roughly speaking, we are looking for an {At }t 0 -adapted function K (·) : [0, T ] → RC 2 (Ω, A; H ) such that, firstly, the composition t → F[t , K (t )] induces a set-valued mapping G K which can be used for nonautonomous stochastic differential inclusions and, secondly, its reachable set ϑG K (t , K (0)) coincides with K (t ) at every time instant t ∈ [0, T ]. Definition 4.1. Let a set-valued map F[t , M ] : {t } × Ω × H ; H × Lin(U , H ) in PLIPΛ ({t } × Ω × H ; H × Lin(U , H )) be given for every t ∈ [0, T ] and M ∈ RC 2 (Ω, At ; H ). The mapping K : [0, T ] → RC 2 (Ω, A; H ) is called a solution to the stochastic morphological evolution equation ◦

“ K (t ) = F[t , K (t )] ” in [0, T ]

(11)

if it satisfies the following conditions: (1) K is adapted w.r.t. { A t }t 0 , i.e., K (t ) ∈ RC 2 (Ω, At ; H ) for any t ∈ [0, T ]. (2) [0, T ] → ERC , t → (t , K (t )) is sequentially continuous with respect to RC . (3) The set-valued composition G K : [0, T ] × Ω × H ; H × Lin(U , H ) given by

(t , ω, x) → F[t , K (t )] (t , ω, x) belongs to PLIPΛ ([0, T ] × Ω × H ; H × Lin(U , H )) in the sense of Definition 3.3. (4) K (t ) = ϑG K (t , K 0 ) for every t ∈ [0, T ]. The term “morphological equation” was introduced by Aubin in [3,4] and, it originally concerns an extension of ordinary differential equations to compact subsets of Rn supplied with the Hausdorff metric. Recently this concept has been extended to random closed subsets of Rn in [21, § 3.7] and the ◦ term “stochastic morphological equation” was coined there for the first time. Here the notation K (t ) is used in a formal way only. Indeed, the stochastic morphological evolution equation (11) is just a symbolic representation of the “integral” set condition (4). Theorem 4.2 (Existence for stochastic morphological evolution equations). Assume that the following properties of the single-valued function F = (F 1 , F 2 ) hold for Λ > 0 and T > 0: (1) For every time t ∈ [0, T ] and every random closed set M ∈ RC 2 (Ω, At ; H ), the value F[t , M ] belongs to PLIPΛ ({t } × Ω × H ; H × Lin(U , H )), i.e., (a) F[t , M ] (t , ω, x) ⊂ H × Lin(U , H ) is nonempty, bounded, closed for all ω ∈ Ω , x ∈ H , (b) F[t , M ] (t , ·, x) is At -measurable for every x ∈ H , (c) F[t , M ] (t , ω, ·) : H ; H × Lin(U , H ) is Λ-Lipschitz for every t ∈ [0, T ] and ω ∈ Ω , Def.

(d) F[t , M ] lg = sup{F[t , M ] (t , ω, 0)∞ | ω ∈ Ω} < ∞. .  > 0 such that sup{F[t , M ] lg | t ∈ [0, T ], M ∈ RC 2 (Ω, At ; H )}  γ (2) There exists γ (3) F[t , M ] is locally Lipschitz continuous in the sets and continuous in time in the following sense: For any radius R > 0, there exists a constant λ0, R > 0 and a modulus of continuity ω0, R (·) such that



2

dl∞ F[t1 , M 1 ] (t 1 , ·,·), F[t2 , M 2 ] (t 2 , ·,·)

     λ0, R · RC (t 1 , M 1 ), (t 2 , M 2 ) + ω0, R |t 1 − t 2 |

for any (t 1 , M 1 ), (t 2 , M 2 ) ∈ ERC with E( M j 2∞ )  R 2 .

2964

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Then at every square integrably bounded random closed set K 0 ∈ RC 2 (Ω, A0 ; H ), there starts a unique solution K : [0, T ] → RC 2 (Ω, A; H ) of the stochastic morphological evolution equation (11) with square integrably bounded values. Proof. As in the proof of [17, Theorem 3.2] for nonlocal single-valued stochastic differential equations, the proof here uses interpolated Euler-like approximations on equi-distant partitions of [0, T ], which turn out to be a uniform Cauchy sequence with respect to dlRC . The completeness of RC 2 (Ω, At ; H ) for each t ∈ [0, T ] (see Lemma 2.9) then provides us with a candidate K : [0, T ] → RC 2 (Ω, A, P) for the sought solution. To begin, first extend each given set-valued map F[t , M ] : {t } × Ω × H ; H × Lin(U , H ) to

F[t , M ] : [t , T ] × Ω × H ; H × Lin(U , H ),

(s, ω, x) → F[t , M ] (s, ω, x)

i.e., constantly for the nonlocal dependence at time t. This extension is predictable with respect to A since F[t , M ] (t , ·, x) : Ω ; H × Lin(U , H ) is At -measurable (by assumption (1)(b)) and {At }t 0 is increasing. Hence, F[t , M ] ∈ PLIPΛ ([t , T ] × Ω × H ; H × Lin(U , H )) for any (t , M ) ∈ ERC . For every n ∈ N with 2n > T set hn := 2Tn , tkn := khn for k = 0, . . . , 2n and K n (0) = K 0 . Write

[s]n := sup tkn tkn  s, k ∈ N for any s ∈ [0, T ] and n ∈ N. Then, for each k = 1, . . . , 2n , extend the curve K n : (tkn , tkn+1 ] →

RC 2 (Ω, A; H ) inductively by means of the random closed reachable set of the nonautonomous stochastic differential inclusion







d X t (ω) ∈ A X t (ω) + F[1tn , K (tn )] t , ω, X t (ω) n k k





dt + F[2tn , K (tn )] t , ω, X t (ω) dW t (ω), n k k

i.e., for t ∈ (tkn , tkn+1 ] and k = 1, . . . , 2n ,

K n (t ) = ϑF[tn , K k

 n n (tk )]

 

t ; tkn , K n tkn .

As a consequence of this piecewise definition of K n , the set-valued map G n : [0, T ] × Ω × H ; H × Lin(U , H ) given by

(t , ω, x) → F[[t ]n , K n ([t ]n )] (t , ω, x) belongs to PLIPΛ ([0, T ] × Ω × H ; H × Lin(U , H )) and satisfies

K n (t ) = ϑG n (t ; 0, K 0 )

(12)

for every t ∈ [0, T ]. Inequalities (6), (7) of Proposition 3.4 imply for every 0  s  t  T and n ∈ N that

2       2t , E  K n (t )∞  e C E  K 0 2∞ + 2γ          RC s, K n (s) , t , K n (t )  (t − s) · C E  K 0 2∞ + γ

(13)

with a constant C = C (Λ, ρ , T ) < ∞. In particular, all these sets K n (t ) ∈ RC (Ω, At ; H ) satisfy the following a priori estimates:

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2965



2 T < ∞, (a) E( K n (t )2∞ )  R 2 with the radius R := e C /2 E( K 0 2∞ ) + 2γ



2

dlRC K m (t ), K n (t )

(b)

 2   2   2 · dlRC K m (t ), S (t ) K 0 + dlRC S (t ) K 0 , K n (t )         2 · RC (0, K 0 ), t , K m (t ) + RC (0, K 0 ), t , K n (t )     .  const(Λ, ρ , T ) · E  K 0 2∞ + γ

(13)

Now choose the Lipschitz constant λ0, R > 0 and a nondecreasing modulus of continuity ω0, R (·) related to R according to assumption (3). The comparative inequality (9) in Proposition 3.4 then guarantees for all indices n > m and at each time t ∈ [0, T ] that

dlRC



2 K m (t ), K n (t ) 

t



2

sup dlRC F[[s]m , K m ([s]m )] (s, Y ), F[[s]n , K n ([s]n )] (s, Y )

0

t 

Y ∈L2



ds · C e Ct

       λ0, R · RC [s]m , K m [s]m , [s]n , K n [s]n + ω0, R (hm ) ds · C e Ct .

0

In particular, triangle-like inequality (B.3) provides for n > m and s ∈ [0, t ]

      RC [s]m , K m [s]m , [s]n , K n [s]n               2 e 2ρ T · RC [s]m , K m [s]m , [s]n , K m [s]n + RC [s]n , K m [s]n , [s]n , K n [s]n           2    + dlRC K m [s]n , K n [s]n  2 e 2ρ T · [s]n − [s]m C E  K 0 2∞ + γ       2   + sup dlRC K m (·), K n (·) .  2 e 2ρ T · hm C E  K 0 2∞ + γ

(13)

[0, s ]

Hence for any n > m, the auxiliary function δm,n : [0, T ] → [0, ∞[, t → sup[0,t ] dlRC ( K m , K n )2 is bounded, monotone increasing and satisfies for every t ∈ [0, T ]

t



 

 



 + 1 · hm + ω0, R (hm ) . 2λ0, R C e C T · δm,n (s) ds + const(Λ, ρ , T ) · E  K 0 2∞ + γ

δm,n (t )  0

Gronwall’s inequality holds even for monotone integrable functions such as δm,n [21, Proposition A.1] and so, we obtain the uniform bound for any n > m

 



 



 + 1 · hm + ω0, R (hm ) . sup dlRC ( K m , K n )2  const(λ0, R , Λ, ρ , T ) · E  K 0 2∞ + γ

[0, T ]

For each time instant t ∈ [0, T ], the sequence ( K n (t ))n∈N is a Cauchy sequence in the complete metric space (RC (Ω, At ; H ), dlRC ). Denote its limit by K (t ) ∈ RC (Ω, At ; H ). Furthermore, the Cauchy property is even uniform with respect to t ∈ [0, T ], from which follows the uniform convergence





sup dlRC K n (t ), K (t ) → 0 for n → ∞.

t ∈[0, T ]

2966

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Furthermore [0, T ] → ERC , t → (t , K (t )) is sequentially continuous w.r.t. RC (although RC is not a metric), i.e. for any sequence (tk )k∈N converging to some t ∈ [0, T ],

    RC tk , K (tk ) , t , K (t ) → 0 for k → ∞. Indeed, triangle-like inequality (B.3) and uniform estimate (13) imply for any 0  s < t  T

             RC s, K (s) , t , K (t )  inf 8e 6ρ T RC s, K (s) , s, K n (s) + RC s, K n (s) , t , K n (t ) n∈N

    + RC t , K n (t ) , t , K (t )     . = 8e 6ρ T · (t − s) · C E  K 0 2∞ + γ Hence, K satisfies conditions (1) and (2) defining a solution of a stochastic morphological evolution equation in Definition 4.1. The same arguments imply the uniform convergence





 



sup RC [t ]n , K n [t ]n , t , K (t )

t ∈[0, T ]

→ 0 for n → ∞.

From assumption (3) on the continuity of F , the set-valued mappings G n : [0, T ] × Ω × H ; H × Lin(U , H ), given by

G n (t , ω, x) := F[[t ]n , K n ([t ]n )] (t , ω, x),

n ∈ N,

converge uniformly to the mapping G : [0, T ] × Ω × H ; H × Lin(U , H ) given by (t , ω, x) → F[t , K (t )] (t , ω, x). Hence, G ∈ PLIPΛ ([0, T ] × Ω × H ; H × Lin(U , H )) (in the sense of Definition 3.3). Finally, it follows from Eq. (12) and inequality (9) in Proposition 3.4 that









0 = lim dlRC ϑG (t ; 0, K 0 ), ϑG n (t ; 0, K 0 ) = lim dlRC ϑG (t ; 0, K 0 ), K n (t ) n→∞

n→∞

at every time instant t ∈ [0, T ], i.e., K (t ) = ϑG (t ; 0, K 0 ).

2

The proof reveals that the assumption E( K 0 2∞ ) < ∞ can be replaced by a stronger condition on F[t , M ] by means of inequality (7) in Proposition 3.4: Corollary 4.3. In addition to assumptions (1) of Theorem 4.2, suppose: (2 ) All the values of F are uniformly bounded in the sense that







sup F[t , M ] (t , ω, x) H ×Lin(U , H ) t ∈ [0, T ], M ∈ RC 2 (Ω, At ; H ),



ω ∈ Ω, x ∈ H < ∞.

(3 ) F[t , M ] is locally Lipschitz continuous in the sets and continuous in time in the following sense: For any set M 0 ∈ RC 2 (Ω, A0 ; H ) and radius R > 0, there exist a constant λ0, R > 0 and a modulus of continuity ω0, R (·) such that



2

dl∞ F[t1 , M 1 ] (t 1 , ·,·), F[t2 , M 2 ] (t 2 , ·,·)

     λ0, R · RC (t 1 , M 1 ), (t 2 , M 2 ) + ω0, R |t 1 − t 2 |

for any (t 1 , M 1 ), (t 2 , M 2 ) ∈ ERC with RC ((t 0 , M 0 ), (t j , M j ))  R. Then every random closed set K 0 ∈ RC 2 (Ω, A0 ; H ) (not necessarily square integrably bounded) initialises a unique solution K : [0, T ] → RC 2 (Ω, A; H ) of the stochastic morphological evolution equation (11).

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2967

5. Reachable random sets as dynamical systems The solution of a nonautonomous evolution equation depends on both the actual time t and the initial time t 0 and not just on the elapsed time t − t 0 as in an autonomous system. The solution mapping φ(t , t 0 , x0 ) of an initial value problem for which an existence and uniqueness theorem holds then satisfies the initial value property

φ(t 0 , t 0 , x0 ) = x0 and the two-parameter semigroup evolution property

  φ(t 2 , t 0 , x0 ) = φ t 2 , t 1 , φ(t 1 , t 0 , x0 ) ,

t0  t1  t2 ,

as well as the continuity property

(t , t 0 , x0 ) → φ(t , t 0 , x0 ) is continuous on an appropriate state space X . Dafermos [12] and Hale [14] called an abstract nonautonomous dynamical systems with such properties a process. This term was used in a deterministic context and is not to be confused with a stochastic process. Nevertheless, the reachable set mappings (introduced above for stochastic differential inclusions) and stochastic morphological evolution equations also define a deterministic two-parameter semigroup or process generalised to allow for time-dependent domains X t := RC 2 (Ω, At ; H ), i.e., with

φ(t , t 0 , ·) : RC 2 (Ω, At0 ; H ) → RC 2 (Ω, At ; H ),

t  t0 ,

where φ(t , t 0 , K 0 ) := ϑ F (t ; t 0 , K 0 ) = K (t ) (writing K 0 now instead of x0 ). The above properties follow in this setting from the existence and uniqueness theorems and the Da Prato and Frankowska theorem presented above. In this setting, the dynamical system is deterministic with stochasticity built into or hidden in the time-dependent state spaces. This contrasts with the skew-product like formalism of random dynamical systems considered in Arnold [2]. Moreover, it allows one to use concepts from the theory of nonautonomous dynamical systems such as nonautonomous pullback attractors (see [18]) for stochastic differential inclusions and stochastic morphological evolution equations. Furthermore it provides a mean-square analysis rather than pathwise as in [2], but allows the dynamics to depend nonlocally on the different sample paths such as on the expectation as in [17]. These topics will be pursued elsewhere. 6. Nucleation and birth-and-growth processes The evolution of random closed sets along stochastic differential inclusions covers a quite broad class of growth processes. In particular, it is not restricted to convex-valued or expanding set evolutions in RC 2 (Ω, A; H ). Theorem 4.2, however, does not consider any form of nucleation. Roughly speaking, the growth process is initiated completely by the initial random closed set K 0 ∈ RC 2 (Ω, A; H ) at an initial time t 0 = 0 and not by some additional random process starting possibly elsewhere and later as in a nucleation process. In this section, we suggest how this restriction can be overcome by means of approximation. Obviously, there is no significant difficulty in solving the following problem in a piecewise way, i.e., for the time of additional nucleation t 1 ∈ (0, T ), the random closed sets K 0 ∈ RC 2 (Ω, A0 ; H ), N 1 ∈ RC 2 (Ω, At1 ; H ) and the same function F = (F1 , F2 ) as given in Theorem 4.2

⎧ ◦ ⎪ ⎨ “ K (t ) = F[t , K (t )] ” in [0, T ], K (0) = K 0 , ⎪ ⎩ K (t 1 ) = lim K (t ) ∪ N 1 . t ↑t 1

2968

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Here the limit on the right-hand side is understood in RC 2 (Ω, At1 ; H ) with respect to dlRC and, its existence results from the Lipschitz continuity of solutions with respect to dl2RC (due to Proposition 3.4 (1)). Theorem 4.2 can now be applied piecewise, i.e., first to [0, t 1 ] and then to [t 1 , T ]. It leads to the unique solution K (·) : [0, T ] → RC 2 (Ω, H ):

 K (t ) =

ϑF [·, K ] (t ; 0, K 0 )

if 0  t < t 1 ,

ϑF [·, K ] (t ; 0, K 0 ) ∪ ϑF [·, K ] (t ; t 1 , N 1 ) if t 1  t  T .

This simple example is based on a single additional nucleation at time t 1 ∈ (0, T ). Now the central question is which type of convergence of the time-dependent nucleation is appropriate for extending this piecewise construction approximatively. Proposition 6.1 (A priori estimate for countably many nucleation processes). Suppose for Λ, T > 0 fixed and the single-valued functions F k = (F1k , F2k ) with k = 1 and 2 that the hypotheses of Theorem 4.2 hold. Let

I 1 , I 2 ⊂ [0, T ] be (at most) countable and contain 0. Assume for the functions N 1 : I 1 → RC 2 (Ω, A; H ) and N 2 : I 2 → RC 2 (Ω, A; H ): (i) N j is adapted w.r.t. {At }t 0 , i.e., N j (t ) ∈ RC 2 (Ω, At ; H ) for every t ∈ I j , j = 1, 2. (ii) N 1 and N 2 are uniformly square integrably bounded, i.e.,



 





 

2 2 η := max sup E  N 1 (t )∞ , sup E  N 2 (t )∞ t∈I1

t∈I2

< ∞.

(iii) The modified Hausdorff excesses between their graphs are bounded in time in the sense that

δ1 (t ) := δ2 (t ) :=

inf

s∈ I 2 ∩[0,t ]

inf

s∈ I 1 ∩[0,t ]



2   |t − s| + e ⊂ , RC N 1 (t ), S (t − s) N 2 (s)

t ∈ I1,



 2  |t − s| + e ⊂ , RC N 2 (t ), S (t − s) N 1 (s)

t ∈ I2,

are bounded. Suppose also that K 1 , K 2 : [0, T ] → RC 2 (Ω, A; H ) satisfy for every t ∈ [0, T ]

K k (t ) =

 s∈ I k ∩[0,t ]

ϑF k



[·, K k ]



t ; s, N k (s) ∈ RC 2 (Ω, At ; H ),

k = 1, 2.

, η, ρ , T ) > 0 such that Then, there is a constant  C = C (Λ, γ



2

dlRC K 1 (t ), K 2 (t )



 C e Ct ·

!



sup δk | I k ∩[0,t ] + t · sup dl∞ F 1 , F 2

2 "

k=1,2

holds for every t ∈ [0, T ]. Proof. The main tools are the explicit continuity estimates of random closed reachable sets with respect to time and initial set as established in Proposition 3.4. The a priori estimate (6) of Proposition 3.4 guarantees E( K k (t )2∞ )  R 2 for all t ∈ [0, T ] and



2 T . k = 1, 2 with the radius R := e const(Λ,ρ , T ) η2 + 2γ Fix ε > 0 arbitrarily. For each s ∈ I 1 , there exists a point of time r ∈ I 2 ∩ [0, s] such that

 2 |s − r | + e ⊂ RC N 1 (s), S (s − r ) N 2 (r )  δ1 (s) + ε

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2969

by definition of δ1 (s). Moreover, inequality (7) of Proposition 3.4 states that

       2 RC r , K 2 (r ) , s, ϑF 2 [·, K 2 (·)] s; r , K 2 (r )  (s − r ) · const(Λ, ρ , T ) η2 + γ and, the inclusions N 2 (r ) ⊂ K 2 (r ) and ϑF 2 [·, K 2 (·)] (s; r , K 2 (r )) ⊂ K 2 (s) hold P-almost surely in Ω due to the characterising assumption about K 2 . Thus,



   2 2  ⊂  2 e ⊂ RC N 1 (s), S (s − r ) N 2 (r ) + e RC S (s − r ) N 2 (r ), K 2 (s)   , η, ρ , T ) δ1 (s) + ε .  const(Λ, γ

2

e ⊂ RC N 1 (s), K 2 (s)

The trajectory-based proof of inequality (9) of Proposition 3.4 and the assumed Λ-Lipschitz continuity of all set-valued maps F[1s, K (s)] and F[2s, K (s)] on H ensure that 1





e ⊂ RC ϑF 1

[·, K 1 (·)]

 C

2





t ; s, N 1 (s) , ϑF 2

[·, K 2 (·)]

 2 e ⊂ RC N 1 (s), K 2 (s)

t +

2

t ; s, K 2 (s)



2 dl∞ F[1s , K 1 (s )] , F[2s , K 2 (s )] ds



s

 C

 2 e ⊂ RC N 1 (s), K 2 (s)

t +





1

 2 2





1

 2 2

2 sup dl∞ F , F

+ λ · RC



 (·, K 1 ), (·, K 2 ) ds



s

 C

 2 e ⊂ RC N 1 (s), K 2 (s)

t +

2 sup dl∞ F , F

2



+ λ · dlRC ( K 1 , K 2 ) ds



s

for every t ∈ [s, T ] with a constant C  1 depending only on Λ, ρ , T . Here λ > 0 denotes the uniform Lipschitz constant of F[1s, K (s)] (s, ·) relative to the radius R as was specified in assumption (3) of the 1 Existence Theorem 4.2. , η, ρ , T )  1 can be speciC = C (Λ, γ Since ε > 0 can be taken arbitrarily small, a larger constant  fied such that







2

e ⊂ RC ϑF 1 [·, K 1 (·)] t ; s, N 1 (s) , K 2 (t )

    2  e ⊂ RC ϑF 1 [·, K 1 (·)] t ; s, N 1 (s) , ϑF 2 [·, K 2 (·)] t ; s, K 2 (s) 



 C · δ1 (s) + (t − s) · sup dl∞ F , F 1

 2 2

t +λ·



 

 2

dlRC K 1 s , K 2 s

 ds



s

for every s ∈ I 1 and t ∈ [s, T ]. Clearly, the excess of a countable union from a fixed set in RC 2 (Ω, At ; H ) is bounded by the supremum of the excesses. Hence,

 2 e ⊂ RC K 1 (t ), K 2 (t )

= 

e ⊂ RC

#

sup

 s∈ I 1 ∩[0,t ]

s∈ I 1 ∩[0,t ]



ϑF 1 [·, K 1 (·)] 

$  t ; s, N 1 (s) , K 2 (t )







e ⊂ RC ϑF 1 [·, K 1 (·)] t ; s, N 1 (s) , K 2 (t )

2970

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

  C·

sup

s∈ I 1 ∩[0,t ]

  C·

sup

s∈ I 1 ∩[0,t ]



1

δ1 (s) + (t − s) · sup dl∞ F , F

 2 2



t 2

+λ·

dlRC ( K 1 , K 2 ) ds



s

 2 δ1 (s) + t · sup dl∞ F 1 , F 2 + λ ·

t



 

 2

dlRC K 1 s , K 2 s

 ds



0

is satisfied for every t ∈ [0, T ]. The same arguments lead to

2



e ⊂ RC K 2 (t ), K 1 (t )

  C·

sup

s∈ I 2 ∩[0,t ]



1

δ2 (s) + t · sup dl∞ F , F

 2 2

t +λ·



 

 2

dlRC K 1 s , K 2 s

 ds



.

0

Finally, the claim about



 2 Def. 2 ⊂  2 = max e ⊂ RC K 1 (t ), K 2 (t ) , e RC K 2 (t ), K 1 (t )

dlRC K 1 (t ), K 2 (t )

results from Gronwall’s inequality applied to the (possibly not continuous, but) bounded and monotone increasing auxiliary function t → sup[0,t ] dlRC ( K 1 , K 2 )2 (as in the proof of Theorem 4.2). 2 This previous proposition clarifies which “distance” between nucleation rules is relevant for the corresponding solutions. So far the nucleation is prescribed by means of a predictable function from I → RC 2 (Ω, A; H ) with an at most countable domain I ⊂ [0, T ]. This result is the main tool for solving a birth-and-growth problem approximatively if the domain I of the nucleation function is possibly not countable. The characterising comparison (14) is formulated via dlRC because it is not immediately clear whether the union for all s ∈ I ∩ [0, t ] is a random closed set (see Remark 6.3). Theorem 6.2 (Existence of solutions to some birth-and-growth problems). As in the Existence Theorem 4.2, assume that Λ > 0, T > 0 and the single-valued function F = (F 1 , F 2 ) satisfy the following properties: (1) The value F[t , M ] belongs to PLIPΛ ({t } × Ω × H ; H × Lin(U , H )) for every time t ∈ [0, T ] and random closed set M ∈ RC 2 (Ω, At ; H ), i.e., (a) F[t , M ] (t , ω, x) ⊂ H × Lin(U , H ) is nonempty and closed for all ω ∈ Ω and x ∈ H , (b) F[t , M ] (t , ·, x) is At -measurable for every x ∈ H , (c) F[t , M ] (t , ω, ·) : H ; H × Lin(U , H ) is Λ-Lipschitz for every t ∈ [0, T ] and ω ∈ Ω , Def.

(d) F[t , M ] lg = sup{F[t , M ] (t , ω, 0)∞ | ω ∈ Ω} < ∞. .  > 0 such that sup{F[t , M ] lg | t ∈ [0, T ], M ∈ RC 2 (Ω, At ; H )}  γ (2) There exists γ (3) F is locally Lipschitz continuous w.r.t. sets and continuous in time in the sense that for every radius R > 0 there exist a constant λ0, R > 0 and a modulus of continuity ω0, R (·) such that



2

dl∞ F[t1 , M 1 ] (t 1 , ·,·), F[t2 , M 2 ] (t 2 , ·,·)

  2   λ0, R · RC (t 1 , M 1 ), (t 2 , M 2 ) + ω0, R |t 1 − t 2 |

for any t j ∈ [0, T ] and M j ∈ RC 2 (Ω, At j ; H ) with E( M j 2∞ )  R 2 for j = 1 and 2.

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2971

Let I ⊂ [0, T ] contain 0 and assume for the function N : I → RC 2 (Ω, A; H ): (4) N (·) is adapted w.r.t. {At }t 0 , i.e., N (t ) ∈ RC 2 (Ω, At ; H ) for every t ∈ I . (5) N (·) is uniformly square integrably bounded, i.e., supt ∈ I E( N (t )2∞ ) < ∞. (6) There exists a sequence (sn )n∈N in I with s1 = 0 and

sup



inf

t ∈ I ∩[0, T ] m∈{1...n}: sm t

 2  |t − sm | + e ⊂ → 0 for n → ∞. RC N (t ), S (t − sm ) N (sm )

Then there exists an {At }t 0 -adapted mapping K : [0, T ] → RC 2 (Ω, A; H ) satisfying for every t ∈ [0, T ]

#



dlRC K (t ),

$   ϑF [·, K (·)] t ; s, N (s) = 0.

(14)

s∈ I ∩[0,t ]

Proof. For each n ∈ N set I n := {s1 . . . sn } ⊂ I , and the Existence Theorem 4.2 (on page 2963) provides a solution K n : [0, T ] → RC 2 (Ω, A; H ) to the problem

K n (t ) =



  ϑF [·, K n (·)] t ; s, N (s) ∈ RC 2 (Ω, At ; H ),

t ∈ [0, T ],

s∈ I n ∩[0,t ]

in a piecewise way as described at the beginning of this Section 6. Proposition 6.1 and the asymptotic assumption (6) about (sn )n∈N in I imply the Cauchy property of ( K n (·))n∈N in the sense that



sup

2

→ 0 for m → ∞.

sup dlRC K n1 (t ), K n2 (t )

n1 ,n2 m t ∈[0, T ]

Hence, the completeness of the metric space (RC 2 (Ω, At ; H ), dlRC ) (see Lemma 2.9) guarantees a limit function K : [0, T ] → RC 2 (Ω, A; H ) which is adapted to {At }t 0 and satisfies



2

dlRC K n (t ), K (t )

→ 0 for n → ∞ and each t ∈ [0, T ].

The asymptotic hypothesis (6) about (sn )n∈N is uniform w.r.t. t ∈ [0, T ], so it follows from the details of the proof of Proposition 6.1 that

e ⊂ RC

#



  ϑF[·, K (·)] t ; s, N (s) ,

s∈ I ∩[0,t ]

$   ϑF[·, K (·)] t ; s, N (s) → 0

 s∈ I n ∩[0,t ]

for n → ∞ and each t ∈ [0, T ]. Obviously, the second argument is contained in the first union, so the corresponding excess is 0. Furthermore, inequality (9) of Proposition 3.4 and the local Lipschitz continuity of F[s,·] in assumption (3) lead to the convergence

# dlRC

 s∈ I n ∩[0,t ]

  ϑF[·, K (·)] t ; s, N (s) ,



$   ϑF [·, K n (·)] t , s, N (s) → 0

s∈ I n ∩[0,t ]

for n → ∞ and every t ∈ [0, T ] by means of Lebesgue’s theorem of dominated convergence. Finally the claimed identity results from the triangle inequality for dlRC . 2

2972

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Remark 6.3. The union



  ϑF [·, K (·)] t ; s, N (s) : Ω ; H

s∈ I ∩[0,t ]

consists of possibly uncountably many sets in RC 2 (Ω, At ; H ) and, it is not obvious if it also belongs to RC 2 (Ω, At ; H ). Each of these random closed reachable sets, however, is constructed by means of strong solutions X t ∈ L 2 (Ω, At , P; H ) to a stochastic differential inclusion in a time interval [s, t ] and, each set of these L 2 functions is decomposable as mentioned in Proposition 3.2. 2 Corollary 2.5 states the existence of M t ∈ RC 2 (Ω, At ; H ) such that SRC ( M t ) coincides with the L 2 closure of the union of all solutions X t ∈ L 2 (Ω, At , P; H ) inducing ϑF [·, K (·)] (t ; s, N (s)) for any initial time s ∈ I ∩ [0, t ]. In this selection-wise sense, M t can be regarded as the closure of the union above:



  ϑF [·, K (·)] t ; s, N (s) = M t ∈ RC 2 (Ω, At ; H ).

s∈ I ∩[0,t ]

Finally, the characterising condition (14) on K : [0, T ] → RC 2 (Ω, A; H ) in Theorem 6.2 is equivalent to K (t ) = M t for every t ∈ [0, T ]. Remark 6.4 (Extension to possibly empty sets of nucleation). An assumption of Theorem 6.2, namely N (t ) ∈ RC 2 (Ω, At ; H ) for each t ∈ I , implies that the subsets N (t )(ω) ⊂ H are nonempty for all t ∈ I and ω ∈ Ω . It serves essentially the rather technical purpose that the mean square Hausdorff excess 2 e ⊂ RC ( N (t ), N (sm )) is well defined in [0, ∞] via L selections of N (t ). For many applications in modelling, however, this hypothesis is not a significant obstacle. If all considerations are restricted to random closed sets in a fixed closed subset V  H , for example, then one should assume

F[t , M ] (t , ·,·) = {0} ⊂ H × Lin(U , H ) in Ω × ( H \ V ) for all t ∈ [0, T ] and M ∈ RC 2 (Ω, At ; H ) anyway so that V is invariant, i.e., the admissible solutions to stochastic differential inclusions cannot “leave” V . Now fix an arbitrary point x0 ∈ H \ V and consider N (t ) ∈ RC 2 (Ω, At ; H ) with the additional feature x0 ∈ N (t )(ω) for all ω ∈ Ω . This modification does not have any explicit influence on the evolution of the random closed sets K (t ) ∈ RC 2 (Ω, At ; H ). Corollary 6.5 (Two special cases of birth-and-growth processes). Suppose that ( S (t ))t 0 is uniformly continuous. In addition to the hypotheses (1)–(3) of Theorem 6.2, suppose that the {At }t 0 -adapted function N : [0, T ] → RC 2 (Ω, A; H ) satisfies both supt ∈[0, T ] E( N (t )2∞ ) < ∞ and one of the following conditions: (a) N is continuous with respect to dlRC , or (b) N is measurable w.r.t. dlRC , right-continuous and expanding in the sense that N (t 1 , ω) ⊂ N (t 2 , ω) P-almost surely in Ω whenever 0  t 1  t 2  T . Then, there exists a function K : [0, T ] → RC 2 (Ω, A; H ) with

K (t ) =



  ϑF [·, K (·)] t ; s, N (s) ∈ RC 2 (Ω, At ; H )

s∈ I ∩[0,t ]

(in the sense of Remark 6.3) for each time instant t ∈ [0, T ].

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Proof. Assuming ( S (t ))t 0 is uniformly continuous, there is a constant

2973

γ > 0 such that

     S (t )x − S (s)x  |t − s|e γ |t −s| ·  S (s)x  |t − s|e (γ +ρ )T x H H H holds for all x ∈ H and 0  s  t  T . This implies the corresponding estimates for the semigroup ( S (t ))t 0 on L 2 (Ω, A, P; H ) (induced via composition) and any X ∈ L 2 (Ω, A, P; H ). By means of L 2 selections, one obtains the uniform estimate





dlRC S (t ) M , S (s) M  |t − s|e (γ +ρ ) T

  E  M 2∞

for any square integrably bounded set M ∈ RC 2 (Ω, Rn ) and s, t ∈ [0, T ]. As a consequence, RC and the metric | · − · | + dlRC (·,·) induce the same concept of sequential convergence in the sense that equivalence (B.4) holds for any (t , M ) ∈ ERC with E( M 2∞ ) < ∞ (even locally uniformly). (a) If N is continuous with respect to dlRC , then it is uniformly continuous in the compact interval [0, T ]. This implies directly the assumptions (4)–(6) of Theorem 6.2 concerning N (·) and a sequence (sn )n∈N . (b) Now let N : [0, T ] → RC 2 (Ω, A; H ) be {At }t 0 -adapted, measurable, right-continuous w.r.t. dlRC and expanding. Due to E( N ( T )2∞ ) < ∞ and the expanding property, N (·) has bounded variation w.r.t. the metric dlRC . According to [11, Theorem 4.3 (b)], there is a set S 0 ⊂ [0, T ] of (at most) countably many points in which N is not continuous. Furthermore, for every s ∈ S 0 ∩ ]0, T ], [11, Lemma 4.1] guarantees some set N − (s) ∈ RC 2 (Ω, A; H ) with dlRC ( N (t ), N − (s)) → 0 for t ↑ s. For s ∈ [0, T ] \ S 0 , set N − (s) := N (s). Then we conclude from the assumption that N (·) is right-continuous: For every ε > 0 and s ∈ [0, T ], there exists a radius δ = δ(ε , s) ∈ ]0, ε [ such that

⎧ ⎨ dlRC  N (t ), N − (s) < ε ⎩

dlRC



for all t ∈ [0, T ], s − δ  t < s, 2  N (t ), N (s) < ε for all t ∈ [0, T ], s  t  s + δ.

In particular, the compact interval [0, T ] ⊂ R can be covered by finitely many open balls B δ(ε,s j ) (s j ), j = 1 . . . n. Set additionally sn+ j := max{0, s j − δ(ε , s j )} for j = 1 . . . n. Then every t ∈ [0, T ] satisfies

inf

j ∈{1 ... 2n}: s j t



 2  |t − s j | + e ⊂ < 2ε . RC N (t ), N (s j )

Indeed, there are only two cases possible: Firstly, if we can choose j ∈ {1 . . . n} with s j  t  s j + δ(ε , s j ) then t − s j  δ(ε , s j ) < ε and dlRC ( N (t ), N (s j )) < ε . Secondly, if there exists an index j ∈ {1 . . . n} with s j − δ(ε , s j )  t < s j , then we have sn+ j  t  sn+ j + ε and













dlRC N (t ), N (sn+ j )  dlRC N (t ), N − (s j ) + dlRC N − (s j ), N (sn+ j ) < 2 ·

ε 2

.

Finally, replace the fixed parameter ε > 0 by a monotonically decreasing sequence εm → 0 and continue this supplementary selection of finitely many points of time in [0, T ] inductively. This procedure provides a (not necessarily monotone) sequence (sn )n∈N in [0, T ] so that

sup

inf

t ∈ I ∩[0, T ] m∈{1...n}: sm t



 2  |t − sm | + e ⊂ → 0 for n → ∞. RC N (t ), N (sm )

2

The severe restriction about uniform continuity of the semigroup ( S (t ))t 0 can be replaced by the assumption about the {At }t 0 -adapted function N : [0, T ] → RC 2 (Ω, A; H ) that the set of all its L 2  2 selections, i.e., t ∈[0, T ] SRC ( N (t )) is precompact in L 2 (Ω, A; H ).

2974

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Appendix A. Filippov-like theorem of Da Prato and Frankowska The focus of interest is an existence theorem for stochastic differential inclusions in the separable Hilbert space H . In particular, a priori estimates that compare a given curve with a solution to the inclusions are needed. Moreover, they should have a form similar to Filippov’s theorem about deterministic differential inclusions (see [5, Theorem 5.3.1] or [21, Theorem A.6]). In 1994, Da Prato and Frankowska presented such an existence result for stochastic differential inclusions with globally Lipschitz continuous drift and diffusion terms [13]. Their main statements even concern Itô integral inclusions with a strongly continuous semigroup on a separable Hilbert space. A year later, Motyl [27] published independent existence and uniqueness results about stochastic differential inclusions in the Euclidean space under some assumptions of dissipative type (see also [23, 24,26] and references therein). Now consider only the “evolutionary” case (with a ρ -contractive semigroup) under the general assumptions of Section 2 and prove such a Filippov-like theorem essentially by means of the arguments of Aubin, Da Prato and Frankowska in [6, Theorem 4.1]. The comparative estimate, however, is slightly modified so that it can be used more easily for verifying the main statements and thus, the proof is presented in a self-contained way here. Theorem A.1 (Filippov-like theorem of Da Prato and Frankowska). Let A be the generator of a ρ -contractive F = strongly continuous semigroup S (·) of linear bounded operators on H . Suppose for the set-valued map  ( F1,  F 2 ) : [0, T ] × Ω × H ; H × Lin(U , H ) that (i) (ii) (iii) (iv)

 F has nonempty bounded and closed values,  F (·,·, x) : [0, T ] × Ω ; H × Lin(U , H ) is predictable for every x ∈ H , F (t , ω, ·) is Λ-Lipschitz for each t ∈ [0, T ], ω ∈ Ω , there is Λ > 0 such that  there is γ > 0 such that  F (t , ω, 0)∞  γ holds for all t ∈ [0, T ], ω ∈ Ω .

Furthermore, let Y := I0 (t 0 , Y 0 , γ0 , σ0 ) be any Itô process with drift γ0 ∈ L2A ([t 0 , T ], H ) and diffusion σ0 ∈ L2A ([t 0 , T ], Lin(U , H )). Then, for every initial random variable X 0 ∈ L 2 (Ω, At0 , P; H ) there exist a drift γ ∈ L2A ([t 0 , T ], H ) and a diffusion coefficient σ ∈ L2A ([t 0 , T ], Lin(U , H )) such that the related Itô process X := I0 (t 0 , X 0 , γ , σ ) satisfies







γ (t , ω) ∈  F 1 t , ω, X t (ω) ,



σ (t , ω) ∈  F 2 t , ω, X t (ω)

for (L1 × P)-almost all (t , ω) ∈ [t 0 , T ] × Ω and

 2



 X − Y I,[t0 ,t ]  C · E | X 0 − Y 0 |

2



t +

  2  C (1+t −t 0 )  E dist γ0 (s), σ0 (s) , F (s, ·, Y s ) ds · e (t −t0 )·e 



t0

for each t ∈ [t 0 , T ] with a constant C  1 that depends only on Λ and ρ . Proof. The proof is based on essentially the same iterative construction of approximate solutions as Theorem 4.1 of [6]. There exist functions γ1 ∈ L2A ([t 0 , T ], H ) and σ1 ∈ L2A ([t 0 , T ], Lin(U , H )) satisfying γ1 (t , ω) ∈  F 1 (t , ω, Y t (ω)) and σ1 (t , ω) ∈  F 2 (t , ω, Y t (ω)) with

   γ0 (t , ω) − γ1 (t , ω) = dist γ0 (t , ω),  F 1 t , ω, Y t (ω) ,      σ0 (t , ω) − σ1 (t , ω) = dist σ0 (t , ω),  F 2 t , ω, Y t (ω) Q

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2975

for every (t , ω) ∈ [t 0 , T ] × Ω as a consequence of Theorem 8.2.11 of [7] about measurable marginal maps and Proposition 2.7. Then, the Itô process X 1 := I0 (t 0 , X 0 , γ1 , σ1 ) satisfies

 1   X − Y 2



I,[t 0 ,t ]

= E | X0 − Y 0|

2



 t



 t

2

+E

|γ1 − γ0 | ds + E

σ1 − σ

t0



= E | X0 − Y 0|

2



 t +E

 2 0Q

ds

t0



2 dist (γ0 , σ0 ),  F (s, Y s ) ds



t0

at every time t ∈ [t 0 , T ]. Now iterate this construction to obtain three sequences (γn )n∈N , (σn )n∈N and ( X n )n∈N in L2A ([t 0 , T ], H ), L2A ([t 0 , T ], Lin(U , H )) and L2A ([t 0 , T ], H ), respectively, with







γn+1 (t , ω) ∈  F 1 t , ω, X tn (ω) ,



σn+1 (t , ω) ∈  F 2 t , ω, X tn (ω)

and

   γn (t , ω) − γn+1 (t , ω) = dist γn (t , ω),  F 1 t , ω, X tn (ω) ,      σn (t , ω) − σn+1 (t , ω) = dist σn (t , ω),  F 2 t , ω, X tn (ω) Q for all (t , ω) ∈ [t 0 , T ] × Ω as well as

X n+1 = I0 (t 0 , X 0 , γn+1 , σn+1 ). Then the uniform Λ-Lipschitz continuity of F (t , ω, ·) and the general estimate (2) for Itô processes imply

 n +1 2 X − Xn

I,[t 0 ,t ]

 t



 t

2

=E

|γn+1 − γn | ds + E t0

 t =E

 2 nQ

σn+1 − σ t0



2   dist γn (s, ·), σn (s, ·) ,  F s, ·, X sn ds



t0

 t E

    2 dl  F s, ·, X sn−1 ,  F s, ·, X sn ds



t0

 t 2

Λ ·E

n −1 2 X − X n ds s



s

t0

 Λ2 · 4e (1+2ρ )(t −t0 ) ·

t

 n   X − X n−1 2

I,[t 0 ,s]

t0

By induction in n it follows for every n ∈ N and t ∈ [t 0 , T ] that

ds.

ds

2976

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

 n +1 2 X − Xn



 2 2+(1+2ρ )(t −t 0 ) n

 Λ e

I,[t 0 ,t ]

t

sn

·

dsn

I,[t 0 ,s1 ]

t0

n   Λ2 e 2+(1+2ρ )(t −t0 ) ·

s2  2 dsn−1 . . .  X 1 − Y 

t0

t

 1   X − Y 2

I,[t 0 ,sn ]

t0

 

1 n! 1 n!

ds1

t0

(t − sn )n−1 dsn (n − 1)!

2 n  Λ2 e 2+(1+2ρ )(t −t0 ) (t − t 0 ) ·  X 1 − Y I,[t 2 2n  Λe 1+(1+ρ )(t −t0 ) ·  X 1 − Y I,[t

0 ,t ]

0 ,t ]

due to the fact that

T  t 0

(t − s)n−1 g (s) ds dt = (n − 1)!

0

T g (s) 0

( T − s)n ds n!

for every Lebesgue-integrable function g : [0, T ] → R and each n ∈ N; see [3, Lemma 1.4.3] or [6, Lemma 4.2]. The series

c (t ) :=

∞ n 1  √ Λe (1+ρ )(1+t −t0 ) n! n =0

is absolutely convergent for every t  t 0 as the d’Alembert ratio test reveals. Hence, ( X n )n∈N is a Cauchy sequence with respect to  · I,[t0 ,t ] for each t ∈ [t 0 , T ], so there exist limits γ ∈ L2A ([t 0 , T ], H ), σ ∈ L2A ([t 0 , T ], Lin(U , H )) and X ∈ L2A ([t 0 , T ], H ) of the sequences (γn )n∈N , (σn )n∈N , ( X n )n∈N , respectively, with X = I0 (t 0 , X 0 , γ , σ ). Furthermore, γ (t , ω) ∈  F 1 (t , ω, X t (ω)) and σ (t , ω) ∈  F 2 (t , ω, X t (ω)) for (L1 ×P)-almost all (t , ω) ∈ [t 0 , T ] × Ω result from the facts that some subsequences of (γn )n∈N , (σn )n∈N , ( X n )n∈N converge to F (t , ω, ·) is continuous by their respective limits pointwise almost everywhere in [t 0 , T ] × Ω and that  assumption (iii). Then, X satisfies

 X − Y I,[t0 ,t ] 

∞  n +1  X − Xn n =1

I,[t 0 ,t ]

  +  X 1 − Y I,[t

0 ,t ]



∞    (Λe 1+(1+ρ )(t −t0 ) )n  X1 − Y  +  X 1 − Y I,[t ,t ] √ I,[t 0 ,t ] 0 n! n =1

=

∞  (Λe (1+ρ )(1+t −t0 ) )n  X1 − Y  √ I,[t 0 ,t ] n! n =0

at each time t ∈ [t 0 , T ], i.e.,

 2  X − Y 2I,[t0 ,t ]  c (t )2  X 1 − Y I,[t  2

= c (t )



0 ,t ]

E | X0 − Y 0|

2



 t +E t0



2 dist (γ0 , σ0 ),  F (s, Y s ) ds

 .

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Finally, in regard to the claimed estimate, it is necessary to verify for some constant that

c (t )2  γ · e (t −t0 )·e

γ (1+t −t 0 )

2977

γ = γ (Λ, ρ ) > 0

.

Due to absolute convergence, c (·) is analytic in [t 0 , ∞). ∞ 1 √ Λn · n(1 + ρ )en(1+ρ )(1+t −t0 ) dt n! n =1 ∞ % n Λn en(1+ρ )(1+t −t0 ) = Λ(1 + ρ )e (1+ρ )(1+t −t0 ) + (1 + ρ ) · (n − 1)! n =2 & ∞ 2 (1+ρ )(1+t −t 0 ) Λn en(1+ρ )(1+t −t0 )  Λ(1 + ρ )e + (1 + ρ ) · (n − 2)! n =2 √ (1+ρ )(1+t −t 0 ) = Λ(1 + ρ )e + (1 + ρ ) · 2Λ2 e 2(1+ρ )(1+t −t0 ) · c (t )

0

d

c (t ) =

and Gronwall’s inequality imply







c (t )  c (0) + (t − t 0 ) · Λ(1 + ρ )e (1+ρ )(1+t −t0 ) · e (t −t0 )·

2Λ2 (1+ρ )e 2(1+ρ )(1+t −t 0 )

  γ (1+t −t 0 )  c (0) + 1 · e (t −t0 )·γ e for all t  t 0 with some constant

γ = γ (Λ, ρ )  1. 2

Appendix B. The time-oriented distance The time-oriented distance L introduced in Section 2.4 is not a metric in the classical sense, but nevertheless leads to the same concept of sequential convergence as the norm of [0, T ] × L 2 (Ω, A, P; H ). By means of L 2 selections, this concept of time-oriented distance can be extended to random closed sets in RC 2 (Ω, A; H ) (and their tuples with time). Consider the basic set



ERC := (t , M ) t ∈ [0, T ], M ∈ RC 2 (Ω, At ; H ) and define for (t 1 , M 1 ), (t 2 , M 2 ) ∈ ERC , h ∈ [0, T − t 1 ]







2 S (h) M 1 := S (h) f f ∈ SRC ( M 1 ) ∈ RC 2 (Ω, At1 +h , P; H ),

   |t 1 − t 2 | + dlRC ( S (t 2 − t 1 ) M 1 , M 2 )2 if t 1  t 2 , RC (t 1 , M 1 ), (t 2 , M 2 ) := |t 1 − t 2 | + dlRC ( M 1 , S (t 1 − t 2 ) M 2 )2 if t 1 > t 2 .

(B.1)

The representation of dlRC via L 2 selections mentioned in Section 2.2 leads to the relationship

   RC (t 1 , M 1 ), (t 2 , M 2 ) = max

sup

inf

2 f ∈SRC ( M 1 ) g ∈SRC ( M 2 )

sup

2

inf

2 2 g ∈SRC ( M 2 ) f ∈SRC ( M 1 )

  L (t 1 , f ), (t 2 , g ) ,

  L (t 1 , f ), (t 2 , g ) .

(B.2)

2978

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

Finally, the counterpart of inequality (3) (generalising the time-oriented triangle inequality) also holds for tuples in ERC : For any (t 1 , M 1 ), (t 2 , M 2 ), (t 3 , M 3 ) ∈ ERC with t 1  t 2  t 3 ,

     RC (t 1 , M 1 ), (t 3 , M 3 )  2 e 2ρ (t3 −t2 ) · RC (t 1 , M 1 ), (t 2 , M 2 )   + RC (t 2 , M 2 ), (t 3 , M 3 ) .

(B.3)

As a consequence of inequality (B.3), the “open balls backward in time”, i.e.,









B− ρ (t 0 , M 0 ) := (t , M ) ∈ ERC t  t 0 , RC (t 0 , M 0 ), (t , M ) < ρ



for all (t 0 , M 0 ) ∈ ERC and ρ > 0, satisfy four sufficient conditions, which determine a topology on ERC uniquely: For each (t 0 , M 0 ) ∈ ERC consider



N − (t 0 , M 0 ) := U ⊂ ERC ∃ρ > 0: B − ρ (t 0 , M 0 ) ⊂ U . Then the following four implications hold: (a) (b) (c) (d)

U ⊂ U ⊂ ERC and U ∈ N − (t 0' , M 0 ) ⇒ U ∈ N − (t 0 , M 0 ), U1 , . . . , Un ∈ N − (t 0 , M 0 ) ⇒ nj=1 U j ∈ N − (t 0 , M 0 ), U ∈ N − (t 0 , M 0 ) ⇒ (t 0 , M 0 ) ∈ U , U ∈ N − (t 0 , M 0 ) ⇒ ∃V ∈ N − (t 0 , M 0 ): ∀(t , M ) ∈ V : U ∈ N − (t , M ).

− Hence there exists a unique topology T− RC on ERC with respect to which N (t 0 , M 0 ) consists all neighbourhoods of each (t 0 , M 0 ) ∈ ERC according to [9, § I.1, Proposition 2] or [28, Proposition 2.16]. Correspondingly the “open balls forward in time”, i.e.,









B+ ρ (t 0 , M 0 ) := (t , M ) ∈ ERC t  t 0 , RC (t 0 , M 0 ), (t , M ) < ρ



for all (t 0 , M 0 ) ∈ ERC , ρ > 0, induce the neighbourhoods in a (possibly different) topology T+ RC on ERC , which is also uniquely determined. Whenever a curve Γ : [0, T ] → ERC is sequentially continuous with respect to RC in the sense that

∀(tk )k∈N in [0, T ]:

lim tk = t ∈ [0, T ]

k→∞

⇒





lim RC Γ (tk ), Γ (t ) = 0,

k→∞

+ it is continuous with respect to both T− RC and TRC (and vice versa). In this paper, we mostly focus on approximating sequences in ERC and their convergence w.r.t. RC . Hence, the topological aspects in terms of “open subsets” or “open neighbourhoods” are of minor priority here. Equivalence (4) gives the relation between sequential convergence in the vector bundle E L w.r.t. L and w.r.t. the standard norm. So far it is not clear, however, how to extend this equivalence to the set ERC which is lacking any obvious linear structure in its second component. Indeed, the proof of equivalence (4) in [21, § 3.10.1] is based on the (standard) representation of the resolvent operator as Laplace transform. To be more precise, the strongly continuous semigroup ( S (t ))t 0 on H always induces a strongly continuous semigroup on ( L 2 (Ω, A, P; H ),  ·  L 2 ) via composition. The auxiliary distance L provides the additional advantage that all curves t → S (t ) X 0 , X 0 ∈ L 2 , are 1-Lipschitz. This advantage is definitely preserved by ERC , RC and the semigroup on RC 2 (Ω, A; H ) defined by relation (B.1) and, that is what we use for the main results in Theorems 4.2, 6.2. It is not immediately clear, however, if the latter semigroup on RC 2 (Ω, A; H ) is strongly continuous with respect to dlRC , unless one of the following cases is assumed:

P.E. Kloeden, T. Lorenz / J. Differential Equations 251 (2011) 2950–2979

2979

1. the semigroup ( S (t ))t 0 is uniformly continuous, so its generator is a bounded linear operator on L 2 (Ω, A, P; H ), or 2 2. the selection set SRC ( M ) of each random closed set M ∈ RC 2 (Ω, A; H ) under consideration is 2 compact in L (Ω, A; H ). (This quite severe restriction will not be used before the final Corollary 6.5.) In each of these cases one concludes from equivalence (4) by means of L 2 selections that for any bounded sequence ((tk , M k ))k∈N in ERC and (t , M ) ∈ ERC ,





lim RC (tk , M k ), (t , M ) = 0

k→∞

⇐⇒





lim |tk − t | + dlRC ( M k , M )2 = 0.

k→∞

(B.4)

References [1] G. Aletti, E.G. Bongiorno, V. Capasso, Integration in a dynamical stochastic geometric framework, ESAIM Probab. Stat. (2011), doi:10.1051/ps/2010009, in press. [2] L. Arnold, Random Dynamical Systems, Springer-Verlag, 1998. [3] J.-P. Aubin, Mutational and Morphological Analysis: Tools for Shape Evolution and Morphogenesis, Birkhäuser, 1999. [4] J.-P. Aubin, Mutational equations in metric spaces, Set-Valued Anal. 1 (1993) 3–46. [5] J.-P. Aubin, Viability Theory, Birkhäuser, 1991. [6] J.-P. Aubin, G. Da Prato, H. Frankowska, Stochastic invariance for differential inclusions, Set-Valued Anal. 8 (2000) 181–201. [7] J.-P. Aubin, H. Frankowska, Set-Valued Analysis, Birkhäuser, 1990. [8] G. Beer, Topologies on Closed and Closed Convex Sets, Kluwer, 1993. [9] N. Bourbaki, Eléments de Mathématique, Topologie Générale, chapitres 1 à 4, Springer, 2007. [10] C. Castaing, M. Valadier, Convex Analysis and Measurable Multifunctions, Lecture Notes in Math., vol. 550, Springer, 1977. [11] V.V. Chistyakov, On the theory of multivalued mappings of bounded variation of one real variable, Sb. Math. 189 (5–6) (1998) 797–819. [12] C.M. Dafermos, An invariance principle for compact processes, J. Differential Equations 9 (1971) 239–252. [13] G. Da Prato, H. Frankowska, A stochastic Filippov theorem, Stoch. Anal. Appl. 12 (1994) 409–426. [14] J. Hale, Asymptotic Behavior of Dissipative Systems, Math. Surveys Monogr., Amer. Math. Soc., Providence, 1988. [15] P.E. Kloeden, On general semi-dynamical systems, Abstract of PhD thesis, Bull. Aust. Math. Soc. 13 (1975) 151–154. [16] P.E. Kloeden, General control systems, in: W.A. Coppel (Ed.), Mathematical Control Theory, in: Lecture Notes in Math., vol. 680, Springer, 1978, pp. 119–138. [17] P.E. Kloeden, Th. Lorenz, Stochastic differential equations with nonlocal sample dependence, Stoch. Anal. Appl. 28 (2010) 937–945. [18] P.E. Kloeden, M. Rasmussen, Nonautonomous Dynamical Systems, Math. Surveys Monogr., Amer. Math. Soc., Providence, 2011, in press. [19] Jun Gang Li, Shoumei Li, Yukio Ogura, Strong solution of Itô type set-valued stochastic differential equation, Acta Math. Sin. (Engl. Ser.) 26 (2010) 1739–1748. [20] Th. Lorenz, Morphological control problems with state constraints, SIAM J. Control Optim. 48 (2010) 5510–5546. [21] Th. Lorenz, Mutational Analysis: A Joint Framework for Cauchy Problems in and Beyond Vector Spaces, Lecture Notes in Math., vol. 1996, Springer, 2010. [22] M. Malinowski, M. Michta, Stochastic set differential equations, Nonlinear Anal. 72 (2010) 1247–1256. [23] M. Michta, On solutions to stochastic differential inclusions, Discrete Contin. Dyn. Syst. (Suppl.) (2003) 618–622. [24] M. Michta, Optimal solutions to stochastic differential inclusions, Appl. Math. (Warsaw) 29 (2002) 387–398. [25] I. Molchanov, Theory of Random Sets, Springer, 2005. [26] J. Motyl, Viable solutions of set-valued stochastic equation, Optimization 48 (2000) 157–176. [27] J. Motyl, On the solution of stochastic differential inclusion, J. Math. Anal. Appl. 192 (1995) 117–132. [28] B.V. Querenburg, Mengentheoretische Topologie, second ed., Springer, 1979.